Remote Assessment of Parking Areas for PV Canopies with Deep Image Segmentation and Minimum Bounding Rectangle Polygonization

Parking areas offer great potential for placement of photovoltaic systems. These areas can be identified and assessed for technical solar potential at large scale from remote imagery. We created a deep image segmentation model (U-Net) and trained it using human labeled parking areas within high-definition aerial images covering six square miles. The output probabilities from the segmentation model were converted to a binary classification raster. Contiguous regions classified as parking were polygonized with minimum bounding rectangles. Accuracy of multiple classification thresholds and minimum parking area rectangle sizes were evaluated using the Intersection over Union (IoU) object detection accuracy metric. Using leave-one-out cross-validation to evaluate the deep image segmentation model and polygonization we achieved an IoU of 0.643 and a spatial false positive rate of 2.9%.